Video summary

I Tested 6 Code Graph Tools—Here Is What Actually Happened

Main summary

Key takeaways

Technology

Summary

The video compares six code-graph tools with plain Claude Code on the same substantial redesign of The Gray Cut, a macOS video editor. The task involved rebuilding the start screen and settings, and updating the editor interface to match supplied design files.

Test setup and how the tools work

All seven coding sessions used Opus 5.5 at medium effort, with no subagents. Each started from an identical project copy and received the same prompt and design files. The runs were organized with Fable; setup and discarded attempts were excluded from the reported results.

The tools aim to help Claude find relevant code and understand relationships between parts of a codebase:

  • CodeGraph maps function calls and references, updates the map as code changes, and exposes it through MCP. In this run, its prompt hook supplied graph information without Claude calling its MCP tools.
  • Graphify groups connected code into related areas. Claude queried it through terminal commands, with hooks prompting it to consult the graph.
  • Graft maps function calls and generates code-reference files. This run used MCP; hooks also supplied context about the possible effects of edits.
  • GitNexus traces connections and larger application flows. Claude used its MCP tools to explore the impact of changes.
  • Codebase Memory MCP provides a locally stored, searchable map that Claude can query through MCP.
  • Code Review Graph maps functions and their connections. It also has a separate review workflow, though this run mainly used it to understand the code before editing.

Results

CodeGraph had the lowest reported cost, with plain Claude Code second. CodeGraph used about 17.6 million tokens, versus 19.5 million for plain Claude Code—roughly 10% fewer. At the API prices used for comparison, that was $8.66 versus $9.18, a difference of about 6%, or 52 cents. These are API-equivalent figures; the sessions were run through the creator’s subscription. The middle four tools were within roughly 35 cents of one another, and the creator cautions against reading much into rankings from one run per setup.

Graft was the notable outlier: it used about 39% more tokens and cost about 23% more than plain Claude Code, despite reporting an estimated 486,000 tokens saved. That estimate compared its answers with reading whole files, rather than measuring how much Claude actually avoided reading. Graft also made 92 model requests, compared with 72 for plain Claude Code. The test does not establish why it used more.

Review of the implementations

All seven versions looked broadly close to the intended design. However, Astra’s review found bugs in every implementation, even though each build and regression tests passed.

The CodeGraph version correctly restored saved before-and-after video durations when reopening a project, which five other versions did not. It still had an issue where some controls disappeared after editing or undoing an edit. The creator plans to fix those findings and run the version through their usual quality checks before using it.

Completion times and conclusions

CodeGraph finished in about 19 minutes; GitNexus was close behind. Plain Claude Code took about 23.5 minutes, and Graft nearly 26 minutes. Because the runs overlapped in different groups, the test does not prove that CodeGraph is consistently faster.

The creator’s pick for this task was CodeGraph, based on its cost and preferred implementation, while noting the remaining bugs. Plain Claude Code also performed well, leading the creator to question whether extra code-graph tools reliably improve results with newer models. The video suggests that older or cheaper models might benefit more, but does not test that here.

Main speakers and sources

  • The Gray Cat video creator/narrator — describes the test setup, tool behavior, results, and conclusions.
  • Astra — reviewed the implementations and identified bugs.
  • Claude Code with Opus 5.5 — performed the seven coding sessions.
  • The code-graph tools tested — CodeGraph, Graphify, Graft, GitNexus, Codebase Memory MCP, and Code Review Graph.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video